Long AI tasks can involve repeated waits between actions. On November 2, 2026, Anthropic introduced message threads in public beta to reduce that delay in extended agent sessions. The feature targets responsiveness while leaving token usage unchanged. Anthropic’s announcement explains the launch.
For businesses, the useful question is whether those shorter waits improve a complete workflow. A coding assistant that responds sooner can still spend most of its time running tests, waiting for external tools or producing changes that need review.
Why long agent sessions accumulate overhead
Consider an assistant investigating a software bug. It reads files, examines an error, proposes a change, runs a test and studies the result. Each step adds information that may matter to the next decision.
In a conventional Messages API exchange, the application resends the conversation history. A thread instead retains it temporarily on the server: the first request supplies the full context, and subsequent requests send new messages plus the previous response ID. The API reconstructs the conversation before calling the model. See the message threads documentation.
An everyday analogy is a meeting where participants can continue from the previous discussion instead of receiving the complete briefing again before every question. The topic remains just as complex; less material needs to be transferred between turns.
What the reported 35 percent improvement measures
Anthropic reports that, with threads enabled by default for most Claude Code users, requests containing 500,000 to one million tokens show approximately 35% lower median time to first token. This is a company-reported observation for that context range, rather than an independent benchmark. Source
Time to first token measures the wait before a response begins. It does not establish how quickly an entire assignment finishes.
Imagine a hypothetical task taking ten minutes, of which two minutes consist of waits affected by this optimisation. Reducing those waits by 35% saves 42 seconds, bringing the total to nine minutes and 18 seconds. That would be a 7% improvement overall.
This example is illustrative, not a measured Claude result. It shows why teams should assess elapsed task time alongside response speed.
Smaller requests and lower bills are separate questions
Anthropic says Claude still reads the full conversation on each turn, so threads do not reduce token usage. Standard token rates and existing rate limits continue to apply. Source
Prompt caching addresses a different part of the problem. It allows eligible repeated prompt prefixes to be reused, with separate cache-read and cache-write pricing. Anthropic documents how to enable it with cache_control. Prompt caching documentation
| Mechanism | Main purpose | What to evaluate |
|---|---|---|
| Message threads | Reduce repeated conversation uploads and associated request overhead | Response latency and session continuity |
| Prompt caching | Reuse eligible prompt prefixes | Cache hits, write costs and actual billing |
| Selecting relevant context | Keep the task focused on useful information | Answer quality and missing evidence |
The practical lesson is to budget for completed work. An interface that feels faster and an API bill that becomes smaller are both useful outcomes, but one does not demonstrate the other.
Readers comparing model costs can also consult our analysis of Claude Sonnet 5.5 and cost per task.
Where a pilot could be worthwhile
A reasonable first pilot is a workflow with frequent consecutive turns and a substantial shared history. For example, a team could compare an agent investigating similar software issues with threads enabled and disabled.
Use the same model, tools and acceptance criteria. Record the time until responses begin, the total completion time, the final cost and whether the proposed fix passes review. Include unsuccessful runs: discarding them would make the apparent productivity gain less representative.
By contrast, an occasional question followed by a long human review may offer little opportunity to benefit from session continuity.
For an introduction to choosing suitable tasks, see what you can actually delegate to AI agents.
Temporary continuity requires a recovery plan
Threads expire five minutes after the last request. Applications must retain their own conversation history so they can resend it when continuation fails. Anthropic says thread contents are encrypted in memory, never written to durable storage and deleted within one hour of last use. The feature is not eligible for Zero Data Retention. Documentation
A product team should therefore test interruptions deliberately. Pause the workflow, resume it and check that the application recovers without losing a tool result or repeating an external action.
That last check matters especially when an agent can change records or trigger operations. Restoring conversational context and deciding whether an action already happened are separate engineering responsibilities.
Availability and practical value
Anthropic lists public-beta availability on the native Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry and Anthropic-managed Claude Models on Google Cloud. Launch announcement
Our assessment is that this is a useful infrastructure improvement for teams already experiencing delays in sustained agent workflows. Its value will depend on how much of their waiting time comes from repeated request preparation.
Start with one representative task and a clear success criterion. Faster responses deserve attention when they help users reach an accurate, reviewable result sooner.
Sources
- Anthropic — Make long-running agents faster with message threads, November 2, 2026.
- Claude Platform documentation — Message threads, consulted November 4, 2026.
- Claude Platform documentation — Prompt caching, consulted November 4, 2026.